Incidentally, and I hope this is old news, is that a thread per connection is terribly wasteful of resources and scales poorly with concurrent connections.  A better architecture is to assign a worker thread from a pool when a user connection initiates an interaction. Better still is the number of work threads active is restricted to the about the number of cores with a mechanism to release another worker thread when another stalls on a lock or I/O.  A modest number of threads (greater than number of cores) is fine but massive number of threads just cause contention that needs to be managed.